Disable gradio SSR so Python serves custom routes directly

Inject OpenAI middleware via gradio create_app patch

gradio launch + OpenAI route via ASGI middleware

FastAPI-primary: OpenAI routes before gradio mount, uvicorn server

Insert OpenAI routes ahead of gradio catch-all

gradio-primary launch + attach OpenAI route to gradio app

FastAPI-only server to avoid port 7860 double-bind

Use AutoTokenizer text-only path (repo lacks multimodal preprocessor)

Gemma 4 E4B ZeroGPU OpenAI-compatible endpoint